Weekly AI search research: why one platform label is not enough

AI-visibility measurement often begins with an overly coarse record: platform name, prompt, and answer. That hides conditions that can change the selected sources. One operator may offer ordinary search, deep research, voice, an app, and several models. Google may also release features through different interfaces and countries at different times. A record should therefore describe the exact run: where and in which mode the answer appeared, which model and device were used, the language, country, and time. Without that context, a weekly comparison mixes product changes with changes to the experiment itself.

Section sources:[1] Perplexity[2] ACL Anthology[3] Google Analytics

Perplexity’s Pro Search description shows that focus, selected model, and dialogue mode can be part of the answer condition. This does not establish a ranking mechanism or show that one mode is better for every brand. It does require honest labelling: Pro Search should not be pooled with ordinary search without a qualification. The run log should preserve mode, selected model, source focus, duration, and whether the dialogue continued. A rerun in a fresh context is a different condition, not a technical duplicate.

Section sources:[1] Perplexity

Studies of generative search show that different systems, and even repeated runs of one system, may select different sources. This does not make every answer random, but one attractive screenshot is still weak evidence. Each prompt group needs several repeats under fixed conditions, with the answers preserved. Do not announce a visibility change when it is indistinguishable from ordinary variation between repeats. Choose the sample size using the number of answers that contain source links and a stopping rule set in advance, rather than one universal prompt count.

Section sources:[2] ACL Anthology

Google Analytics documentation matters when interpreting traffic: visits from AI Overviews and AI Mode may be counted under ordinary organic search rather than a separate AI-services channel. A single report of visits from AI systems is therefore likely incomplete. Store the reported source and channel, first page visited, previous-page address, Telegram campaigns, and direct visits separately. If no previous page is reported, that does not prove an AI origin. The dashboard should show this uncertainty instead of presenting a technical channel label as proof of what influenced the user.

Section sources:[3] Google Analytics

Advertising and an ordinary source link answer different questions even when they appear in the same AI interface. A public report should record the interface, date, query set, and identification rule for ads separately. Advertising must not be merged with the share of answers containing a source link or with one blended score: buying an impression does not prove source selection, while an ordinary link says nothing about ad performance. If one domain appears both in an ad and among the sources, those are two observations with different meanings and total counts.

A future study should record the system, exact interface and mode, model, device, language, region, prompt, repeat number, and date for every run. Preserve the full answer and all of its sources separately. Then apply clear rules to record whether the brand was named, a source link was shown, a recommendation was made, an ad appeared, or a user visit followed. Keep a history of corrections. We do not have internal logs from closed systems, so their source selection remains unknown; research results belong to their own samples, and Google Analytics behavior depends on a site's configuration. The defensible conclusion is to improve the method, not publish a universal platform ranking.

Section sources:[1] Perplexity[2] ACL Anthology[3] Google Analytics

Each issue should preserve not only its final thesis but also the record of how that thesis was formed. Useful fields include check date, source version, exact observation, editorial classification, and the reason for a change from the previous week. If a source disappears, changes text, or changes rollout, the old state should not be silently overwritten. Preserve the public URL and describe what changed. This record lets readers distinguish new information from a new interpretation and reduces the risk of turning a note into automated freshness theatre.

Section sources:[1] Perplexity[2] ACL Anthology[3] Google Analytics

To compare weekly results, every record should begin with a run record. It should state the exact prompt, interface language, country, device, place where the check was performed, mode, selected model, and repeat number. Preserve the complete answer with its original links, then mark events separately: source found, object named, link shown, and recommendation made. If the interface changed or part of the answer is unavailable, record that as a limitation instead of filling the gap with a guess. This record is more useful to an editor than one overall score because it shows which claims can be published, which need another check, and which must remain hypotheses. Several records from a week form a reproducible dataset without mixing different products under one platform name.

Section sources:[1] Perplexity[2] ACL Anthology[3] Google Analytics

For replication, read this issue as a measurement protocol rather than a platform ranking. Before comparing runs, fix the launch time, model, interface or mode, device, and query set, then preserve the original answer and its links. Record separately what was observed and what the editorial team infers. If the mode or region changes, create a new observation row instead of merging it with the earlier run. This lets an editor explain why two similar checks diverged and which differences belong to the environment rather than to page quality.